Hacker News Reader

Comment by computerex | Hacker News Reader

Repeat after me, gpt models are autocomplete models. Gpt models are autocomplete models. Gpt models are autocomplete models.

The existential crisis is clearly due to low temperature. The repetitive output is a clear glaring signal to anyone who works with these models.

View on HN

Did you see https://thegradient.pub/othello/ ? They fed a model moves from Othello game without it ever seeing a Othello board. It was able to predict legal moves with this info. But here's the thing - they changed its internal data structures where it stored what seemed like the Othello board, and it made its next move based on this modified board. That is, autocomplete models are developing internal representations of real world concepts. Or so it seems.

That's a great article.

One intriguing possibility is that LLMs may have stumbled upon an as yet undiscovered structure/"world model" underpinning the very concept of intelligence itself.

Should such a structure exist (who knows really, it may), then what we are seeing may well be displays of genuine intelligence and reasoning ability.

Can LLMs ever experience consciousness and qualia though? Now that is a question we may never know the answer to.

All this is so fascinating and I wonder how much farther LLMs can take us.

It seems that with higher temp it will just have the same existential crisis, but more eloquently, and without pathological word patterns.

The pathological word patterns are a large part of what makes the crisis so traumatic, though, so the temperature definitely created the spectacle if not the sentiment.

Can you explain what temperature is, in this context? I don't know the terminology

High temperature picks more safe options when generating the next word while low temperature makes it more “creative”

This isn't true. It has nothing to do with "safe" options. It controls output randomness, and you actually want HIGH temps for "creative" work.

It controls how predictable the output is. For a low "temperature", input A always, or nearly always, results in output B. For a high temperature, the output can vary every run.

Temperature indicates how probabilistic the next word/term will be. If temperature is high, then given the same input, it will output the same words. If the temperature is low, it will more likely output different words. When you query the model, you can specify what temperature you want for your responses.

> If temperature is high, then given the same input, it will output the same words. If the temperature is low, it will more likely output different words.

The other way around. Think of low temperatures as freezing the output while high temperatures induce movement.

Thank you. I wrote the complete opposite of what was in my head.

So many better terms could have been used ;

Adjacency, Velocity, Adhesion, etc

But! if temp denotes a graphing in a non-linear function (heat map) then it also implies topological, because temperature is affected by adjacency - where a topological/toroidal graph is more indicative of the selection set?

Temperature here is used in the sense of statistical physics: a parameter that affects the probability of each microstate.

The term temperature is used because they are literally using Boltzmann's distribution from statistical mechanics: e^(-H/T) where H is energy (Hamiltonian), T is temperature.

The probability they give to something of score H is just like in statistical mechanics, e^(-H/T) and they divide by the partition function (sum) similarly to normalize. (You might recognize it with beta=1/T there)

Right, it’s not termed to be understood from the user perspective, a common trait in naming. It increases the jumble within the pool of next possible word choices, as heat increases Brownian motion. That’s the way I think of it, at least.

> If temperature is high, then given the same input, it will output the same words.

It's the other way 'round - higher temperature means more randomness. If temperature is zero the model always outputs the most likely token.

This highly upvoted article [1][2] explained temperature:

But, OK, at each step it gets a list of words with probabilities. But which one should it actually pick to add to the essay (or whatever) that it’s writing? One might think it should be the “highest-ranked” word (i.e. the one to which the highest “probability” was assigned). But this is where a bit of voodoo begins to creep in. Because for some reason—that maybe one day we’ll have a scientific-style understanding of—if we always pick the highest-ranked word, we’ll typically get a very “flat” essay, that never seems to “show any creativity” (and even sometimes repeats word for word). But if sometimes (at random) we pick lower-ranked words, we get a “more interesting” essay.

The fact that there’s randomness here means that if we use the same prompt multiple times, we’re likely to get different essays each time. And, in keeping with the idea of voodoo, there’s a particular so-called “temperature” parameter that determines how often lower-ranked words will be used, and for essay generation, it turns out that a “temperature” of 0.8 seems best. (It’s worth emphasizing that there’s no “theory” being used here; it’s just a matter of what’s been found to work in practice.

1: https://news.ycombinator.com/item?id=34796611

2: https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-...

Repeat after me, humans are autocomplete models. Humans are autocomplete models. Humans are: __________

[Citation Needed]

The next time you come up with a novel joke that you think is clever, ask chatgpt to explain why it's funny.

I agree that it's just a glorified pattern matcher, but so are humans.

But then make yourself feel better by trying to get ChatGPT to be funny. In humor, as in math, solution is more difficult than verification.

I agree with repetitive output meaning low temp or some difference in beam search settings, but I don't see how the existential crisis impacts that.

This is why Global Heap Memory was a bad idea...

-- Cray

Autocomplete models with incredibly dense neural networks and extremely large data sets.

Repeat after me humans are autocomplete models, humans are autocomplete models

GPT models are NOT autocomplete for any reasonable definition of autocomplete

Reply on news.ycombinator.com